The House Edge

Appearance

Chapter 4 — Research Note (Phase 1) ​

Status: RESEARCH (Phase 1 — written by agent, awaiting Nick's review) Date: 2026-05-13 Working chapter: 04 — The Searcher

How to read this note ​

This is the source bedrock for the chapter, not prose. Every claim listed below is something the chapter is allowed to state. Anything not listed here requires me to come back and source it before it goes into a draft.

A handful of recent developments — most importantly Solana's April 8, 2026 protocol-level sandwich mitigation and Jupiter's MEV-Protect routing — make some 2024-vintage framings stale. Where that matters, I flag it.

1. Key claims ​

Each numbered claim is something the chapter will state. Every claim is sourced; some are sourced to multiple places where the figure is contested or being triangulated.

Searcher economics — what the role is and how it earns

  1. A searcher is a firm or individual that runs software designed to detect profitable transaction opportunities — arbitrage, liquidations, sandwiches — and to submit transactions that capture them. On a competitive chain, searchers compete with each other primarily by paying priority fees or builder bribes; the winner is the one who pays the most. (Flashbots — MEV and the Limits of Scaling, 2025-06-16; Paradigm — Priority is All You Need, 2024-06)

  2. The competitive equilibrium for searchers in a priority-gas auction is approximately zero net margin: in theory and increasingly in practice, the winning searcher pays the full extractable value as priority fees, capturing only the small wedge between their bid and the runner-up's. (Paradigm — Priority is All You Need, 2024-06)

  3. The searcher business has consolidated. The number of active CEX-DEX searchers on Ethereum fell from 23 in late 2024 to 11 in Q1 2025, with the top three capturing 90% of volume. On Base, two firms account for >80% of arbitrage spam transactions. (arXiv 2507.13023 — Measuring CEX-DEX Extracted Value, 2025-07; Flashbots — MEV and the Limits of Scaling, 2025-06-16)

  4. Searcher costs are dominated by failed bids. On Base, roughly 350 failed transactions are emitted for every successful arbitrage; on Ethereum and Solana the ratio is lower but the principle is the same — gas paid on losing attempts is a constant tax on operating revenue. (Flashbots — MEV and the Limits of Scaling, 2025-06-16)

The three flavors — what searchers actually do

  1. Arbitrage (closing a price gap between two venues) is the morally-neutral case: the searcher's profit is the price gap, paid by whichever venue was slow to reprice. There is no direct counterparty whose execution was worsened. A single Solana arbitrage bot ("2Fast") cleared ~$1.9M in profit on one transaction in January 2024 by back-running a low-liquidity-pool mistake. (EigenPhi — 2024 in MEV recap, 2025-01)

  2. Liquidations (forcibly closing under-collateralized borrows) are the economically useful case: searchers earn a protocol-defined liquidation bonus (typically 5–15% on Aave V3, market-specific on Morpho; ~5% on MarginFi) in exchange for keeping lending protocols solvent. On March 10, 2026, ~$27M of wstETH-collateralized positions on Aave V3 were liquidated within hours after the protocol's CAPO oracle mispriced wstETH 2.85% below market. (Aave V3 risk parameters; Morpho liquidation docs; CoinDesk — Aave $27M wstETH liquidations, 2026-03-10)

  3. Sandwich attacks (placing a frontrun buy and a backrun sell around a victim's swap) are the pure-extraction case: the searcher's profit is the worsened execution price the victim received. Unlike arbitrage, sandwiches require a visible pending transaction; unlike liquidations, they perform no useful function for any protocol. (Helius — Solana MEV Report, 2025-Q1)

The 2025–2026 picture

  1. Ethereum sandwich extraction fell roughly 4× over 2025: from ~$10M/month in late 2024 to ~$2.5M/month by October 2025, while monthly DEX volume rose from $65B to >$100B. The decline tracks the migration of order flow to solver-mediated venues (CoW, UniswapX) and protected-routing services (MEV-Share, Flashbots Protect). (EigenPhi data via Cointelegraph Research, 2025-12; EigenPhi tweet 2025-12-08)

  2. Solana sandwich activity was concentrated in a small number of programs: one program ("Vpe") executed 1.55M sandwich transactions in 30 days during Dec 2024 – Jan 2025, profiting ~$13.4M, paying ~$4.6M in Jito tips, with an 88.9% success rate. Jito's own analysis attributed roughly half of all Solana sandwich attacks to that single program. (Helius — Solana MEV Report, 2025-Q1)

  3. On April 8, 2026, Solana shipped protocol-level changes (private transaction routing combined with Jito's Block Assembly Marketplace confidential-execution surface) that closed off the simple validator-leader sandwich path. Coverage estimates 2024–2026 cumulative Solana sandwich extraction at $370M–$500M; that aggregate figure is widely circulated but not traceable to a single primary source. (Cryptopolitan — Solana no longer threatened by sandwich attacks, 2026-04-08; Edgen — Solana shuts down sandwich attacks April 8) Flag: this is recent enough that no post-patch measurement has been published. Treat extraction figures as order-of-magnitude.

  4. CEX-DEX latency arbitrage — searchers who exploit price gaps between centralized exchanges and on-chain DEXs — extracted $233.8M from Ethereum traders over 19 months (Aug 2023 – Mar 2025) across 7.2M arbitrages on $241.7B of routed DEX volume. This is "the darkest corner" of MEV because the inputs (CEX prices, searcher inventories) are not on-chain and so cannot be measured by inspecting the chain alone. (arXiv 2507.13023 — Measuring CEX-DEX Extracted Value, 2025-07)

Jupiter and the protected-execution counterfactual (for the worked example)

  1. Jupiter's combination of dynamic slippage, real-time slippage estimation, and MEV-Protect routing (which sends transactions directly to Jito as bundles, bypassing the public path that exposed them to leader-validator sandwiches) reports +0.6 bps positive average slippage and ~34× greater sandwich protection versus competing aggregators. The honest framing for the book: an unprotected $10,000 USDC→SOL swap on a volatile Solana pair could plausibly lose $50–$150 to a sandwich in early 2025; the same swap routed through Jupiter MEV-Protect in 2026 should lose essentially zero. (Jupiter — Dynamic Slippage announcement, 2024-Q4; Jupiter dynamic-slippage-config GitHub)

Architecture comparison — what Hyperliquid does differently

  1. Hyperliquid's order book has no public mempool; matching happens inside HyperBFT with ~0.07s finality and strict price-time priority. There is no surface on which a third-party searcher can observe a pending swap and front-run it. The economic surplus that would otherwise accrue to external searchers on a sandwich-vulnerable architecture instead accrues to the protocol-owned HLP vault (which absorbs liquidation surplus) and to the validator set. (Hyperliquid order book docs; A1 Research — Hyperliquid: The CLOB King, 2025)

JIT liquidity (the asterisk on the sandwich definition)

  1. JIT (just-in-time) liquidity provision — a searcher mints a concentrated LP position immediately before a swap and burns it after — is mechanically distinct from sandwiching: it adds liquidity at execution, gives the trader price improvement bounded by twice the pool fee tier, and extracts its profit from passive LPs rather than from the trader. As of August 2025, ~$750B of $760B (~97%) of JIT liquidity events on Uniswap V3 came from the top 25 wallets. Whether JIT counts as MEV is contested. (Uniswap — JIT Liquidity, 2023; Kaiko — JIT Liquidity on Uniswap V3, 2025-08; arXiv 2509.16157 — Strategic Analysis of JIT, 2025-09)

2. Numbers to verify ​

Every figure the chapter might use, with status flags. Bold = chapter is likely to lean on it. Italic = optional color.

#NumberSourceDateFlag
N1Ethereum sandwich extraction ~$2.5M/month as of Oct 2025; ~$10M/month late 2024Cointelegraph/EigenPhi2025-12Solid primary data
N2~$40M total Ethereum sandwich extraction in 2025EigenPhi tweet2025-12-08Researcher tweet; cite EigenPhi proper
N3"Vpe" Solana sandwich program: $13.43M profit in 30 days, $4.63M Jito tips paid, 88.9% success rateHelius2025-Q1Single-window snapshot; do not annualize without caveat
N4$370M–$500M cumulative Solana sandwich extraction (16-month window ending April 2026)Cryptopolitan2026-04-08Order-of-magnitude only. Number circulates across outlets without a clean primary source. The chapter should either cite a band ("hundreds of millions") or skip the aggregate.
N50.72% of Helius-validated Solana blocks contained at least one sandwich, over 60 daysBlockchain.News citing @ItsDave_ADA2025-10Useful for the "but most blocks are clean" point; individual validators up to 27.34%
N6$233.8M CEX-DEX extraction, Aug 2023 – Mar 2025, on $241.7B routed volumearXiv 2507.130232025-07Cleanest academic source we have
N7Active Ethereum CEX-DEX searchers fell from 23 → 11 in Q1 2025; top 3 capture 90%arXiv 2507.130232025-07Anchors the "consolidation" point
N8Base: 56% of gas consumed by spam bots, 14% of fees paid; 350 failed txns per successful arb; $0.12 profit at $0.02 fees per winFlashbots2025-06Anchors "real business, real costs"
N9Two firms responsible for >80% of arbitrage spam on BaseFlashbots2025-06
N10MEV contributed ~0.36% of total Solana staking rewards in Q1 2026 (down from 2024–early-2025 highs)Figment Q1 2026 Solana Validator Report2026-Q1Figment-published — use carefully given the book's commercial framing
N11Jito tips peaked ~$25.16M on Jan 20, 2025 (TRUMP launch); Jan 2025 Solana REV $551.7MHelius H1 2025 report2025-07Strong color number, but a one-day peak; do not present as steady-state
N12Jared 2.0 / jaredfromsubway.eth cumulative $50M+ Ethereum sandwich profit since 2023Joel Breaking Down B91 case study2024Vintage but useful as the canonical named MEV bot
N132Fast bot earned ~$1.9M in a single Solana txn, January 2024EigenPhi via Yellow.com2024-01Color cameo for arbitrage
N14Aave V3 liquidation bonus: 5–15% depending on collateral asset; default close factor 50% (up to 100% in distressed cases)Aave docsliveCurrent parameter set
N15March 10, 2026: $27M Aave V3 wstETH liquidations after CAPO oracle mispriced wstETH 2.85% below marketCoinDesk2026-03-10Strong recent named event for the liquidations cameo
N16Aave single-day liquidation peaks: ~$231M (Aug 5 2024), $170M (Feb 3 2025), >$250M (Oct 10 2025 macro event), >$400M (Jan 31–Feb 5 2026 cycle peak)Aave historical liquidationsliveUseful to show liquidations are now macro-correlated
N17JIT liquidity on Uniswap V3: $750B of $760B (~97%) from top 25 wallets, YTD through Aug 2025Kaiko2025-08Strongest "JIT is an oligopoly" stat we have
N18JIT erodes passive-LP returns by up to 44% per affected tradearXiv 2509.161572025-09Academic — anchors "JIT extracts from LPs, not traders"
N19Jupiter MEV-Protect + dynamic slippage reports +0.6 bps positive slippage; ~34× claimed sandwich protection vs. competitorsJupiter2024-Q4Jupiter's own number; protected-execution counterfactual for the worked example
N20Hyperliquid: ~0.07s finality, no mempool, on-chain CLOBHyperliquid docsliveArchitecture, not a measurement

Numbers I deliberately did not include:

  • A 2026 YTD "total MEV $" headline by chain — no clean primary source. (See open question #1.)
  • A specific Hyperliquid MEV $ figure — structurally unmeasured. (See open question #2.)
  • A clean "X% of searcher gross revenue goes to priority fees" ratio — widely repeated in commentary but not published with methodology by Flashbots/EigenPhi/Frontier. We have the theoretical result (Paradigm — ceiling = full MEV) and the examples (Flashbots — $0.12/$0.02 on Base). I plan to use those two, not the unsourced ratio.

3. Contested or evolving claims ​

These need to be handled carefully — either flagged in-text, or framed in a way that survives further drift.

  • Solana sandwich status, post-April 8, 2026. Coverage says the simple validator-leader sandwich is mostly dead. There is no rigorous post-patch measurement yet (only ~5 weeks since the patch as of today). The chapter should describe what was extractable rather than asserting it's permanently fixed. The honest line: "As of mid-2026, the simple Solana sandwich attack is mostly suppressed; more sophisticated extraction via CEX-DEX latency and toxic order flow remains live."

  • Does JIT count as MEV? Uniswap's framing is that JIT is competitive liquidity provision that benefits traders; academic and trader-side framing treats it as extraction from passive LPs. The chapter should not settle this — it should describe both framings and let the reader see that the answer depends on whose pocket you're tracking.

  • The "$370M–$500M Solana sandwich" aggregate. Several outlets cite it, none link to a primary source. I will use a hedged phrasing ("by mid-2026 estimates running into the hundreds of millions of dollars cumulatively") and footnote both the Helius point-in-time measurement (clean) and the aggregate (less clean).

  • Searcher consolidation as a permanent feature vs. a phase. The arXiv 2507.13023 paper shows consolidation in the CEX-DEX segment specifically. Whether this generalizes to all searcher activity is unsettled — sandwiches on Solana were also concentrated (Vpe ~half of attacks per Jito), but the long tail of arb on retail-grade L2s looks more competitive. The chapter should be specific about which segment it's describing.

  • MEV's share of validator rewards. Figment's Q1 2026 figure (0.36% of staking rewards) is much lower than 2024 highs (~30%+ in some windows). Depending on which window the chapter cites, the answer to "how big a deal is MEV to validator economics" is very different. Using a 2024 number would be misleading; the Q1 2026 number is current.

4. Characters introduced ​

The Searcher is the chapter's anchor. Everyone else is a cameo whose full treatment is in another chapter.

The Searcher (full sidebar in this chapter)

  • Who they are: A firm or operator running latency-optimized software that monitors mempools, builder feeds, and exchange APIs for profitable transaction opportunities. Most established searchers are small teams (2–15 engineers) running on dedicated infrastructure colocated with relays and builders.
  • How they earn: From the price gap (arbitrage), the protocol-paid bonus (liquidations), or the worsened execution price of a victim (sandwiches). Always settled in the same transaction or bundle that captures the opportunity.
  • How they spend: On gas/priority fees (the dominant cost — paid on every attempt, win or lose), on infrastructure (colocation, RPC nodes, in-house simulators), and on engineering talent. The competitive moat is increasingly an exclusive relationship with a block builder, not raw algorithmic edge.
  • TradFi analogue: A high-frequency trading desk at Jane Street or Citadel Securities. Same role, same incentives, different rails. The honest framing for the reader: what HFT did to NYSE in 2007–2012, searchers did to on-chain markets in 2020–2024.

The validator / builder (cameo here; full sidebar in Chapter 5)

  • The actor who decides transaction order inside a block, and who collects the priority fees / builder bribes the searcher pays.

The unsuspecting trader (the "you" of the worked example)

  • The retail or business user whose pending swap is the input to a sandwich, or whose imprecise pricing creates the gap an arbitrage closes.

The liquidatable borrower (cameo in the liquidations section)

  • A DeFi user whose collateral value fell below the maintenance threshold, exposing them to forced repayment by a searcher.

Named programs / firms appearing once (for color, not full sidebar treatment):

  • jaredfromsubway.eth — canonical Ethereum sandwich bot
  • Vpe — dominant Solana sandwich program in late-2024 / early-2025
  • 2Fast — Solana arbitrage bot (single $1.9M txn)
  • Jupiter — Solana DEX aggregator running MEV-Protect; appears in the protected-execution counterfactual
  • Jito — block engine / auction operator on Solana; one-line cameo, full treatment in Chapters 5/6
  • Aave — lending protocol; appears in the liquidations cameo
  • HLP — Hyperliquid's protocol-owned LP vault; appears in the chain-comparison box

5. Worked example candidates ​

The SPEC's preference was Option 1 (a $10,000 USDC→SOL swap on Jupiter sandwiched for $73, continuing the prologue). The 2026 facts complicate this. I'm presenting three viable framings; see Open Question Q1 — Nick needs to choose.

Candidate A — "The unprotected 2025 sandwich" (closest to SPEC's preference, retrospective framing) ​

A retail trader swaps $10,000 of USDC for SOL on Jupiter in early 2025, before enabling Jupiter's MEV-Protect mode. On a volatile mid-cap pair, the trade is sandwiched for $50–$150 — call it $73 to continue the prologue's example. The chapter walks through the four-panel sandwich diagram (mempool sighting → frontrun → user fills at degraded price → backrun).

  • Pros: continues the prologue cleanly, lets the chapter develop the sandwich mechanic in its purest form, vivid, easy to follow the dollar.
  • Cons: the example is a 2024–early-2025 vintage. After April 2026, this exact attack is largely suppressed. The chapter would need to be explicit that this is a retrospective, ending with a section like "This attack used to net $73; today, on Solana specifically, the same trade routed through Jupiter MEV-Protect would lose closer to zero. The next paragraphs explain why, and where the value went instead."
  • Decision risk: framed correctly, this is the strongest of the three. Framed sloppily, it reads as out-of-date.

Candidate B — "The CEX-DEX latency arb" (current, harder to make vivid) ​

An MM running an Ethereum CEX-DEX strategy spots a 12 bps gap between Binance's ETH/USDT pair and Uniswap V3's ETH/USDC pool. It posts a fill on Binance, picks up ETH cheaply, and dumps it on Uniswap, paying a builder bribe to be included at the top of the next block. Net profit: $1,200. The counterparty on the Uniswap side: a passive LP, who effectively bought the searcher's overpriced ETH.

  • Pros: current as of mid-2026; demonstrates the "darkest corner" where extraction is hardest to see; the loser (the LP) is non-obvious and pedagogically valuable.
  • Cons: harder to make vivid because the loser is a passive LP, not a human-coded "you." The reader doesn't viscerally feel the $1,200 the way they feel a $73 sandwich.

Candidate C — "The wstETH liquidation, March 10, 2026" ​

An Aave V3 borrower has $200,000 of wstETH posted as collateral against a $150,000 ETH debt. On March 10, 2026, Aave's CAPO oracle prices wstETH 2.85% below the secondary market. The borrower's health factor crosses the threshold. Within seconds, a searcher repays $75,000 of the borrow and seizes ~$78,750 of wstETH (5% liquidation bonus), netting $3,750 minus gas — say $3,200. The borrower lost their position not because the market moved but because the oracle was wrong; the searcher captured the bonus regardless.

  • Pros: a real named event, fresh (~2 months old), shows the "useful-ish but uncomfortable" face of MEV.
  • Cons: liquidations are inside-baseball; the cold open and the reader's first-introduction-to-MEV beat both work better with a sandwich. Also: the oracle was wrong, which complicates the "this is just market structure" framing.

Agent's recommendation ​

Candidate A, framed as a retrospective, paired with brief cameos of B (a paragraph) and C (a paragraph using the March 10 wstETH event). The chapter's anchor is the sandwich, because that's the cleanest expression of the adversarial thesis; the cameos give the reader the full spectrum (price discovery, useful, pure extraction) the SPEC's learning objective #2 requires.

The retrospective framing is actually a feature, not a bug: the chapter can end by showing that even the cleanest extraction technique gets architecturally responded to, then asking "where did the value go instead? to the validators, to the builders, to Jito's bundle market" — which is the perfect transition into Chapter 5.

6. Open questions for Nick ​

Q1 — Worked-example framing (most important). The SPEC was written assuming the $73 Solana sandwich is a current phenomenon. As of April 8, 2026, the simple version of that attack is largely suppressed by a protocol-level patch and by Jupiter's MEV-Protect routing. Three options:

(a) Frame Candidate A as a retrospective: "this is what the attack looked like through early 2025, and here's how the structure changed." My recommendation. (b) Move to Candidate B (CEX-DEX arb) as the anchor and demote the sandwich to a multi-paragraph treatment. (c) Find a sandwich that still works in 2026 — likely a long-tail Solana memecoin pair or an L2 — and use that, at the cost of moving away from the "Jupiter major-pair" frame the prologue probably sets up.

Q2 — Figment self-citation. The Q1 2026 Solana validator report (Figment Insights) is the freshest source we have for the "MEV is ~0.36% of validator rewards" number. Given the book's commercial framing (Figment is named only in the epilogue), I'd prefer to cite a non-Figment source for this same figure if one exists. Is there a Helius / Galaxy / Frontier piece I should look harder for, or is the Figment cite fine because the chapter is about the searcher rather than the validator?

Q3 — The $370M–$500M Solana sandwich aggregate. Several outlets cite this band for the 16-month window ending April 2026. I cannot find a primary source. I'll either (a) skip the aggregate and use only the clean Helius monthly snapshot, or (b) hedge ("estimates running into the hundreds of millions cumulatively"). Preference?

Q4 — Searcher gas-spend ratio. The widely-repeated figure "80–90% of gross searcher revenue is paid out as priority fees" is not sourced cleanly. I propose to drop the ratio and instead make the point via Paradigm's theoretical ceiling + Flashbots' Base examples. Is the ratio-less framing acceptable, or do you want me to chase a primary source?

Q5 — JIT as MEV-or-not. The SPEC asks me to "confirm whether the chapter should distinguish 'true' sandwiches from JIT liquidity provision." Yes — they are mechanically and economically distinct, and the literature treats them as separate phenomena. My proposed treatment: a short footnoted aside (~50 words) on JIT in the sandwich section, framing it as "the asterisk on the sandwich definition." Full JIT treatment, if needed, can live in Chapter 2 (Liquidity). Acceptable?

Q6 — Hyperliquid in the chain-comparison box. Hyperliquid's chain-comparison paragraph is going to read structurally different from Solana's and Ethereum's: "this attack doesn't exist here, but here's where the rents went." That's the honest answer. I want to make sure that's the right framing — i.e., the chapter is allowed to say "the architecture doesn't generate this kind of MEV" without it sounding like a Hyperliquid endorsement.

Q7 — TradFi analogue to use. The Bible encourages naming Citadel and Jane Street. The strongest analogue for searcher economics is the HFT-arbitrage desk circa 2010 (the Flash Boys era). Is that the analogue the chapter should foreground, or do you prefer a different one (specialists at the NYSE, payment-for-order-flow at Robinhood)? PFOF feels like a better fit for Chapter 7; Flash-Boys-era HFT feels like the right one here.

Sources cited in this note ​

Primary (in claim order):

News and aggregators (cited because primary not available):

Phase 1 is complete. No prose, outline, or draft will be produced until Nick reviews this note and approves it (with any changes) for Phase 2.